Tomographic Image Estimation Model Training with Virtual Projection Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radiography systems performing tomosynthesis imaging face challenges in generating high-quality tomographic images due to the limited irradiation angle range, which results in insufficient projection images for training neural networks.
Innovation Solution
An image processing device and a learning device that utilize a tomographic image estimation model trained with machine learning on data comprising three-dimensional structures projected onto virtual irradiation positions, allowing for the generation of high-quality tomographic images from a limited set of projection images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is trained using projection images from tomosynthesis imaging, then the model can generate tomographic images, but the limited irradiation angle range results in insufficient training data
Solution Approach 1:
The patent creates virtual copies of projection images by applying synthetic transformations (noise addition, compression artifacts, motion artifacts) to existing projection images. This copying approach generates additional training data from limited original images, resolving the contradiction between limited available images and the need for sufficient training data.
Solution Approach 2:
The patent transforms existing projection images by changing various parameters including adding different types of noise, applying compression at different rates, introducing motion artifacts, and modifying other image properties. These parameter changes create diverse training samples from limited original data, enabling effective model training despite the limited irradiation angle range.
2Quantity of substance
If the irradiation angle range is increased to obtain more projection images, then training data sufficiency improves, but the device complexity and imaging time increase
Solution Approach 1:
Rather than physically increasing the irradiation angle range to capture more images, the patent synthesizes additional training images through digital copying and transformation of existing projection images. This virtual expansion of data avoids the need for complex hardware modifications or extended imaging procedures.
Solution Approach 2:
The patent replaces the mechanical approach of physically moving the radiation source to additional angles with a computational approach that synthesizes additional training data through image processing and transformation algorithms, eliminating the need for increased device complexity.
3Reliability
If more projection images are captured to improve training data quality, then the model training effectiveness improves, but the imaging time and radiation exposure increase
Solution Approach 1:
The patent performs preliminary transformations and syntheses of training data from existing projection images before model training. By pre-processing and expanding the training dataset through virtual image generation, the system achieves effective training without requiring additional actual imaging time or radiation exposure.
Solution Approach 2:
The patent creates multiple synthetic copies of the limited projection images with various transformations applied, generating sufficient training data without requiring additional physical imaging. This copying strategy maintains training effectiveness while avoiding increased imaging time and radiation exposure.
Data Source
AI summary
An image processing device acquires a plurality of projection images, inputs the acquired plurality of projection images to a tomographic image estimation model, which is a trained model generated by performing machine learning on a machine learning model using learning data composed of a set of correct answer data that is three-dimensional data indicating a three-dimensional structure and of a plurality of virtual projection images, onto which the three-dimensional structure has been projected by performing pseudo-projection on the three-dimensional structure with radiation at a plurality of virtual irradiation positions using the three-dimensional data, and which receives the plurality of projection images as an input and outputs an estimated tomographic image group, and acquires the estimated tomographic image group output from the tomographic image estimation model.